TRAINING A NEURAL NETWORK PROCESSOR FOR DIAGNOSIS OF A CONTROLLED LIQUID CHROMATOGRAPHY PUMP UNIT

Abstract

Training a neural network processor is described for providing diagnostic information of a controlled liquid chromatography pump unit. The training includes executing a sequence of operations wherein the neural network processor is trained with input signals obtained from a simulated version of the controlled liquid chromatography pump unit and associated sensors, while modifying the simulated version of the liquid chromatography pump unit to a pump fault simulation signal. Dependent on a value of the pump fault simulation signal, the simulated version of the liquid chromatography pump unit simulates operation of the liquid chromatography pump unit free from faults or the operation thereof with one or more pump faults. The trained neural network processor obtained therewith is thereafter integrated with a controlled liquid chromatography pump unit to provide for auto-diagnostic capabilities or used in a separate diagnostic unit for diagnosing one or more controlled liquid chromatography pump units not having auto-diagnostic capabilities.

Claims

1. A method of training a neural network processor for providing diagnostic information of a controlled liquid chromatography pump unit comprising executing the following operations: simulating a pump controller generating one or more pump control signals in accordance with a pump controller state; simulating a pump fault simulation signal representative for the presence or absence of a pump fault; executing a simulation model of a pump unit, the simulation model providing one or more simulated sensor signals representing operational parameters of the simulated pump unit in response to the one or more simulated pump control signals, the simulation model selectively simulating a pump fault in response to the pump fault simulation signal; supplying the neural network processor with a combination of input signals comprising the one or more simulated sensor signals and one or more signals indicative for said pump controller state and/or indicative for the supplied one or more simulated pump control signals; computing, by the neural network processor, one or more output signals in response to the supplied combination of input signals; computing a loss function by comparison of a diagnostic state as indicated by the one or more output signals and a diagnostic state as indicated by the pump fault simulation signal; and training the neural network processor by feeding back a loss computed with the loss function, wherein during one or more executions of the operations, the pump fault simulation signal indicates that a predetermined pump fault is to be simulated, and wherein during one or more executions of the operations, the pump fault simulation signal indicates that the predetermined pump fault is to be absent in the simulation.

2. The method according to claim 1, wherein the neural network processor to be trained comprises a combination of: a set of Convolutional Neural Networks (CNN), and a pair of Long Short-Term Memory (LSTM) layers.

3. A method of manufacturing a controlled liquid chromatography pump unit having an auto-diagnostic facility, comprising: providing a liquid chromatography pump unit; providing a pump controller; providing at least one sensor; providing a neural network processor; training the neural network processor using a simulation model for the provided liquid chromatography pump unit, the provided pump controller and the provided at least one sensor, wherein the training comprises the following operations: simulating a pump controller generating one or more pump control signals in accordance with a pump controller state; simulating a pump fault simulation signal representative for the presence or absence of a pump fault; executing a simulation model of a pump unit, the simulation model providing one or more simulated sensor signals representing operational parameters of the simulated pump unit in response to the one or more simulated pump control signals, the simulation model selectively simulating a pump fault in response to the pump fault simulation signal; supplying the neural network processor with a combination of input signals comprising the one or more simulated sensor signals and one or more signals indicative for said pump controller state and/or indicative for the supplied one or more simulated pump control signals; computing, by the neural network processor, one or more output signals in response to the supplied combination of input signals; computing a loss function by comparison of a diagnostic state as indicated by the one or more output signals and a diagnostic state as indicated by the pump fault simulation signal; and training the neural network processor by feeding back a loss computed with the loss function, wherein during one or more executions of the operations, the pump fault simulation signal indicates that a predetermined pump fault is to be simulated, and wherein during one or more executions of the operations, the pump fault simulation signal indicates that the predetermined pump fault is to be absent in the simulation; assembling the provided liquid chromatography pump with the provided pump controller, the provided at least one sensor, and the trained neural network processor to obtain the controlled liquid chromatography pump unit, wherein the assembling comprises: connecting the provided liquid chromatography pump to control outputs of the provided controller for providing control signals to the provided liquid chromatography pump; assembling the provided at least one sensor with a sensor output for providing a sense signal indicative for an operational characteristic of the provided liquid chromatography pump; and connecting the trained neural network processor to one or more of the control outputs of the provided pump controller and to the sensor output of the provided at least one sensor.

4.-7 (canceled)

8. A method of manufacturing a controlled liquid chromatography pump unit according to claim 3 and subsequently using the controlled liquid chromatography pump unit, said subsequently using comprising: providing, by the pump controller, control signals to the liquid chromatography pump; receiving, by the trained neural network processor, respective one or more control signals from the one or more of the control outputs of the pump controller; and receiving, by the trained neural network processor, the sense signal from the output of the at least one sensor, providing, in accordance with the receiving the control signals and receiving the sense signal, one or more diagnostic output signals.

9. A method of manufacturing a system comprising a controlled liquid chromatography pump unit and a diagnostic unit, the method comprising: providing a liquid chromatography pump unit; providing a pump controller; providing at least one sensor; providing a neural network processor; training the neural network processor using a simulation model for the provided liquid chromatography pump unit, the provided pump controller and the provided at least one sensor, wherein the training comprises the following operations: simulating a pump controller generating one or more pump control signals in accordance with a pump controller state; simulating a pump fault simulation signal representative for the presence or absence of a pump fault; executing a simulation model of a pump unit, the simulation model providing one or more simulated sensor signals representing operational parameters of the simulated pump unit in response to the one or more simulated pump control signals, the simulation model selectively simulating a pump fault in response to the pump fault simulation signal; supplying the neural network processor with a combination of input signals comprising the one or more simulated sensor signals and one or more signals indicative for said pump controller state and/or indicative for the supplied one or more simulated pump control signals; computing, by the neural network processor, one or more output signals in response to the supplied combination of input signals; computing a loss function by comparison of a diagnostic state as indicated by the one or more output signals and a diagnostic state as indicated by the pump fault simulation signal; and training the neural network processor by feeding back a loss computed with the loss function, wherein during one or more executions of the operations, the pump fault simulation signal indicates that a predetermined pump fault is to be simulated, and wherein during one or more executions of the operations, the pump fault simulation signal indicates that the predetermined pump fault is to be absent in the simulation; wherein the method of manufacturing further comprises: assembling the provided liquid chromatography pump with the provided pump controller and the provided at least one sensor, therewith: connecting the provided liquid chromatography pump to control outputs of the provided controller for providing control signals to the provided liquid chromatography pump; assembling the provided at least one sensor with a sensor output for providing a sense signal indicative for an operational characteristic of the provided liquid chromatography pump to obtain the controlled liquid chromatography pump unit; providing the controlled liquid chromatography pump unit with a pump interface coupled to the provided pump controller and to the provided at least one sensor; and providing the diagnostic unit with the trained neural network processor, and with the diagnostic unit interface to allow inputs of the trained neural network processor to receive one or more control signals of the provided pump controller and the sense signal of the provided at least one sensor from the pump interface.

10-13. (canceled)

14. A method of manufacturing a system comprising a controlled liquid chromatography pump unit and a diagnostic unit according to claim 9, and subsequently using the system, said use comprising: providing, by the pump controller, control signals to the liquid chromatography pump; receiving, by the trained neural network processor of the diagnostic unit, one or more control signals of the pump controller; receiving, by the trained neural network processor of the diagnostic unit, the sense signal of the at least one sensor obtained by the diagnostic unit interface from the pump interface; and providing, in accordance with the receiving the control signals and receiving the sense signal, one or more diagnostic output signals.

15. The method of claim 14, where the pump interface buffers one or more control signals of the pump controller and the sense signal of the at least one sensor during a first period of time and transfers said buffered signals to the diagnostic unit interface in a second period of time succeeding said first period of time.

16. The method of claim 14, wherein the one or more diagnostic output signals are indicative for a need of maintenance to avoid an expected occurrence of a pump fault in the (near) future.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0032] These and other aspects are described in more detail with reference to the drawing. Therein:

[0033] FIG. 1 schematically shows a first embodiment of a controlled liquid chromatography pump unit having an auto-diagnostic facility;

[0034] FIG. 2 shows various examples of pump faults that may result in a malfunctioning;

[0035] FIG. 3 shows a first embodiment of a system comprising a controlled liquid chromatography pump unit and a diagnostic unit;

[0036] FIG. 4 shows an alternative embodiment of a system comprising a controlled liquid chromatography pump unit and a diagnostic unit;

[0037] FIG. 5 shows aspects of an alternative embodiment of a pump unit;

[0038] FIG. 6, 7A, 7B and 7C schematically show embodiments of a method of training a neural network processor for providing diagnostic information of a controlled liquid chromatography pump unit; therein FIG. 6 shows the method as a sequence of steps to be executed and FIG. 7A, 7B and 7C cluster wise represent these steps for three embodiments;

[0039] FIG. 8A shows an exemplary neural network processor module as employed in FIG. 7A, 7B, 7C in more detail;

[0040] FIG. 8B shows the exemplary neural network processor module upon completion of the training procedure;

[0041] FIG. 9 shows a first example of a neural network processor that can be trained for use in a controlled liquid chromatography pump unit having an auto-diagnostic facility or in a separate diagnostic unit;

[0042] FIG. 9A shows in more detail a component in the exemplary neural network processor of FIG. 9;

[0043] FIG. 10 shows a second example of a neural network processor that can be trained for use in a controlled liquid chromatography pump unit having an auto-diagnostic facility or in a separate diagnostic unit; and

[0044] FIG. 10A shows a component of the exemplary neural network processor of FIG. 10 in more detail.

DETAILED DESCRIPTION

[0045] Like reference symbols in the various drawings indicate like elements unless otherwise indicated.

[0046] FIG. 1 schematically shows a first embodiment of a controlled liquid chromatography pump unit 1 having an auto-diagnostic facility. In the embodiment shown, the pump unit 1 comprises a first controlled pump section 4 and a second controlled pump section 5. The first pump section 4 comprises a first pressure chamber 41 with a first piston 42 driven upwards and downwards by a first motor 44 via a first spindle transmission 43. Similarly, the second pump section 5 comprises a second pressure chamber 51 with a second piston 52 driven upwards and downwards by a second motor 54 via a second spindle transmission 53. A unidirectional fluid path 3 is provided having an inlet 31 and an outlet 39. The unidirectional fluid path 3 has a first fluid path section 3a between a first unidirectional valve 32 and a second unidirectional valve 35 and a second fluid path section 3b between the second unidirectional valve 35 and the outlet 39. The first pump section 4 is connected with the first fluid path section 3a and the second pump section 5 is connected with the second fluid path section 3b. The pump unit 1 further comprises a pump controller 2 that in operation provides drive signals S24, S25 for driving the first controlled pump section 4 and the second controlled pump section 5 in accordance with a controller state. The controlled liquid chromatography pump unit 1 comprises one or more sensors 49, 59, 64, 67 for sensing operational characteristics of the pump unit 1. In the embodiment shown these comprise a first encoder 49 to issue a signal S49 indicative for the position of the first piston 42 and second encoder 59 to issue a signal S59 indicative for the position of the second piston 52 The one or more sensors furthermore comprise a first pressure sensor 64 to issue a sense signal S64 indicative for a fluid pressure in the first fluid path section 3a and a second pressure sensor 67 to issue a sense signal S67 indicative for a fluid pressure in the second fluent path section 3b. As indicated by dashed lines, in some embodiments the controlled liquid chromatography pump unit 1 may comprise one or more flow sensors, for example a first flow sensor 63, configured to issue a sense signal S63 indicative for a flow in the first fluent path section 3a, a second flow sensor 66 to issue a sense signal S66 indicative for a flow through the fluent path from the first pump section 4 to the second pump section 5 and a third flow sensor 68, configured to issue a sense signal S68 indicative for a flow in the second fluent path section 3b. Although flow sensors may provide a valuable indication about the state of the pump, these are preferably avoided to avoid a substantial increase in costs.

[0047] In general the pump controller 2, will use one or more sensors 63, 64, 66, 67, 68 to calculate the drive signals S24, S25. However, the exact operation of the pump controller is not the subject of this disclosure and therefore the sensory inputs of the controller are omitted.

[0048] The controlled liquid chromatography pump unit 1 of FIG. 1 has an auto-diagnostic facility in that it comprises a trained neural network processor 7 having inputs connected to outputs of one or more sensors and to one or more outputs of the pump controller 2. In the embodiment shown, the trained neural network processor 7 has inputs to receive the sense signals S63, S64, S66, S67 and S68 of the sensors 63, 64, 66, 67 and 68 respectively. Also the trained neural network processor 7 has inputs coupled to outputs of the pump controller 2. In this embodiment it receives output signals S247, S257, indicative for the drive signals S24, S25 with which the first and the second pump motor 44, 54 are driven respectively. Furthermore it receives an output signal S27 of the pump controller 2 indicative for the pump controller state. In other embodiments the output signals received by the trained neural network processor 7 from the pump controller 2 may comprise only the output signals S247, S257 indicative for the drive signals S24, S25 or alternatively, may comprise only an output signal S27 of the pump controller 2 indicative for the pump controller state.

[0049] In operation of the controlled liquid chromatography pump unit 1, the pump controller 2 drives the two pistons 42, 52. When one piston presses the fluid into the column, the other piston sucks fluid from the reservoir. In normal circumstances, therewith a desired flow rate can be accurately and precisely maintained, which is essential for a performing a proper chromatography measurement. Various circumstances however may in practice prevent the controlled liquid chromatography pump unit 1 from a proper operation.

[0050] As illustrated in FIG. 2, frequently occurring issues are e.g. a primary seal leakage A, a secondary seal leakage B, an inlet check valve leakage C, an outlet check valve leakage D, a flow path blockage E, a pressure deviation F and a pressure ripple G. In case of a primary seal leakage A, a leakage occurs between first piston 42 and a seal at the inner wall and/or between the seal and the inner wall of the first pressure chamber 41. Analogously, in case of a secondary seal leakage B, a leakage occurs between second piston 52 and a seal at the inner wall and/or between the seal and the inner wall of the second pressure chamber 51. In case of an inlet check valve leakage C fluent leaks back from the first fluent path section 3a towards the inlet. In case of an outlet check valve leakage D fluent leaks back from the second fluent path section 3b towards the first fluent path section 3a. Furthermore, a flow path blockage E may occur, for example the second unidirectional valve 35 or the fluid path from an inlet of the first chamber 41 to the outlet 39 of the pump is blocked or is obstructed by debris. Also it may be the case that a pressure deviation F and/or a pressure ripple G occurs for example because of the second unidirectional valve 35 is partially blocked.

[0051] As noted above a proper operation of the controlled liquid chromatography pump unit 1 is essential to enable reliable chromatography measurements. For this purpose it is important that a failure can be rapidly identified, diagnosed and that the controlled liquid chromatography pump unit can be repaired, preferably by local personnel, so that inactivity is minimized. This is achieved in that the controlled liquid chromatography pump unit 1 has an auto-diagnostic facility including the trained neural network processor 7. In the embodiment shown, the trained neural network processor 7 comprises a respective output for issuing a respective diagnostic output signal 7a, 7b, 7c, 7d, 7e, 7f and 79, which each are indicative for the presence of the pump faults A-G referred to above. The output signals 7a-7g may be of a binary nature, e.g. indicating the diagnosed absence or presence of the corresponding pump fault. Alternatively the output signals may have a higher number of signal levels. For example the signal level may indicate the diagnosed probability or the severity of a fault with a value selected from a value range, wherein the minimum of the range signifies that absence of the fault is diagnosed and the maximum of the range signifies that it is highly likely that the fault is presence or that the fault is severe.

[0052] The presence of the trained neural network renders it possible to provide reliable diagnostics of the controlled liquid chromatography pump unit 1 even in the absence of flow sensors data. Therewith cost of material are modest. Whereas the trained neural network processor 7 in the first place issues a diagnostic output indicative of the nature of a fault, the diagnostic output may further comprise information to assist the local operator to efficiently repair the pump where possible, e.g. by specifying which parts need to be replaced, and which procedural steps are to be taken. If the local operator is not capable to handle a particular fault, the diagnostic information may provide information for a service team.

[0053] Of course, it is even more preferable if the occurrence of a pump fault can be avoided. In some embodiments the diagnostic output facilitates preventive maintenance. Therewith it can be avoided that operation of the pump has to be discontinued at an unfavorable point in time. For example the seal leakage A may be indicated as the amount of fluid leakage in mL/s. With such an output range, the auto-diagnostic facility may also be used for predictive maintenance, i.e. with the amount of fluid leakage and the increase of this amount over time, a proper time for replacement of the piston seal may be suggested. In the example shown, a further output is provided with signal 7o, with which it can be explicitly indicated if the diagnosis reveals that no fault was detected at all. By way of example the signals 7a-7g, 7o may be provided to drive a respective LED indicator at a housing of the controlled liquid chromatography pump unit 1. It is not necessary that the trained neural network processor 7 is permanently active. Maintenance personnel or an operator of the controlled liquid chromatography pump unit 1 may for example deliberately activate the trained neural network processor 7, e.g. by a control button 74 in case a diagnosis is requested. Whereas in this example respective outputs are provided that each provide a diagnostic output signal 7a-7g, 7o, it is alternatively possible that the trained neural network processor 7 is trained to provide a single diagnostic output signal having respective signal levels, each indicative for a particular type of fault.

[0054] As observed above, the one or more diagnostic output signals may be indicative for a need of maintenance to avoid an expected occurrence of a pump fault in the (near) future. In this way it can be avoided that the pump fault actually does occur by a timely maintenance. The pump can continue to function within specifications and its operation has to be interrupted only during the execution of the maintenance.

[0055] It is not necessary that a diagnostic facility is integrated in the controlled liquid chromatography pump unit. FIG. 3 shows an alternative embodiment, wherein the controlled liquid chromatography pump unit 1A is part of a system that further includes a separate diagnostic unit 70. In FIG. 3, elements corresponding to those in FIG. 1 have the same reference. In the example of FIG. 3 the trained neural network processor 7 is replaced with a pump interface 8. The trained neural network processor 7 is incorporated in the separate diagnostic unit 70 which in addition comprises a diagnostic unit interface 71. Furthermore in the example shown, it comprises a display unit 72. This embodiment enables a further cost reduction of the controlled liquid chromatography pump unit 1.

[0056] Whereas FIG. 3 shows a single pair of a controlled liquid chromatography pump unit 1A and a diagnostic unit 70, in practice, the diagnostic unit 70 may be used for diagnosis of a plurality of controlled liquid chromatography pump units 1A. In practice, a controlled liquid chromatography pump unit may operate normally for a long time and it suffices to provide a diagnosis only from time to time, e.g. on a regular basis or upon request. By using therewith a common diagnostic unit 70 the costs of material can be substantially reduced. The double-sided block arrow S87 represents a communication between the pump interface 8 and the diagnostic unit interface 71. The communication typically involves a transmission of the combination of input signals comprising the one or more sensor signals, e.g. S49, S59, S64, and S67 as well as the one or more signals from the pump controller 2. In this case the signals from the pump controller 2 transmitted by the pump interface 8 comprise a signal S28 indicative for the controller state as well as signals S248, S258 indicative for the supplied one or more pump control signals S24, S25. In other examples the signals from the pump controller 2 to the pump interface 8 may comprise a selection of these signals, e.g. only the signal S28 indicative for the controller state or one or more signals S248, S258 indicative for the supplied one or more pump control signals S24, S25. The diagnostic unit interface 71 may also transmit signals to the pump interface 8 for control purposes, e.g. to request a specific selection of the sensor data or data from the pump controller 2. In the exemplary embodiment shown, the pump interface 8 has a storage unit 81 for recording one or more of the combination of input signals S63, S64, S67, S68, S248, S258, and S28 during a period of time. When, in a diagnostic session the diagnostic unit 70 is communicatively coupled to the controlled liquid chromatography pump unit 1A, the stored signals may be transmitted as input for the trained neural network processor 7. Therewith the time for gathering sufficient input data for the trained neural network processor 7 of the diagnostic unit 70 in order to properly diagnose the controlled liquid chromatography pump unit 1 can be reduced. In the diagnostic session the diagnostic unit 70 may transmit control signals to the pump interface 8 to request signal data from the storage unit 81 for a specific period of time. As noted above, communication between the pump interface 8 and the diagnostic unit interface 71 may take place in any of a variety of ways, e.g. wired, wireless, using a private or a public network and the like.

[0057] FIG. 4 shows an alternative embodiment of a controlled liquid chromatography pump unit 1B in a system with a diagnostic unit 70. In this example, the controlled liquid chromatography pump unit 1B comprises in addition to the elements shown in FIG. 2 a section 9 of the pump that comprises a quaternary selection or proportioning valve 91 having an output coupled to the inlet 31 of the fluid path 3 and having a first, a second, a third and a fourth fluid input 92, 93, 94, 95.

[0058] In the examples described with reference to FIGS. 1-4, the controlled liquid chromatography pump unit 1 comprises an arrangement of a first pump section 4 and a second pump section 5 that are arranged in series along a fluid path 3. However, also other arrangements are possible. For example, FIG. 5 shows an alternative embodiment, wherein pump sections 4, 5 are arranged in parallel. The first pump section 4 pumps fluid from a first inlet 31a using a first piston 42 in first pressure chamber 41 having unidirectional valves 32a, 35a. The second pump section 5 pumps fluid from a second inlet 31b using a second piston 52 in second pressure chamber 51 having unidirectional valves 32b, 35b. The pump sections 4, 5 supply the pumped fluid to a common outlet 39. The pistons 42, 52 are driven in counterphase by a pump motor 44′, so that a constant flow is delivered at the common outlet 39. While one of the pump sections 4, 5 fills its pressure chamber with fluid from its inlet, the other one of the pump sections 4, 5 supplies fluid from its pressure chamber to the outlet and reversely. In the example show the pump sections 4, 5 share a common motor 44′. Alternatively, embodiments may be contemplated wherein the pump sections each have a proper motor as long as they drive their respective pump section in counter phase to provide for a constant flow of fluid. FIG. 6, 7A, 7B and 7C schematically show a method of training a neural network processor for providing diagnostic information of a controlled liquid chromatography pump unit comprising an execution of the following sequence of steps S1-S7 as shown in FIG. 6 and as further illustrated for three embodiments in FIG. 7A, 7B and 7C. The drawings in FIG. 7A, 7B, 7C cluster wise represent these steps for illustration purpose. In a practical embodiment of the method, the steps of the method may indeed be performed by mutually different computation devices, but that is not necessary. Instead, the various simulations and computations in this sequence of steps may be performed on a single computation device, e.g. on a general purpose processor.

[0059] Neural networks are trained, using a large set of measurements of pumps operating under normal conditions and with faults such as primary seal leakage, secondary seal leakage B and others. For new pumps, pumps that have been brought to the market recently, such data sets are generally not available. In this disclosure these data sets are generated using simulations.

[0060] In a step S1 a pump controller (e.g. the pump controller 2 of FIG. 1-4) is simulated that generates one or more pump control signals PC in accordance with a pump controller state.

[0061] In step S2 a pump fault simulation signal PFC is simulated that is representative for the presence or absence of a pump fault, for example one of the pump faults A-G as discussed with reference to FIG. 1A.

[0062] In step S3 a simulation model of a pump unit is executed. The simulation model simulates the operation of the pump unit (e.g. the pump unit 3, 4, and 5 in FIG. 1-4 or alternative embodiments thereof as shown in FIG. 5 for example). The simulation model therewith is configured to simulate the operation of the pump in a condition determined by the pump fault simulation signal PFC, in accordance with the pump controller drive signals and controller state and to provide simulated sensor signals representing operational parameters PS that would be obtained from the sensors of the pump unit 3, 4, and 5 operating in that condition.

[0063] Step S1, Step S2 and Step S3 may be sequentially simulated in time instances where the S1 and S2 form the input of S3. The simulated sensor signals and/or other variables of S3 may be fed back into S1 and S2 after which the next sequence starts.

[0064] In step S4 a combination of input signals comprising the one or more simulated sensor signals PS is supplied to the neural network processor. The combination of input signals further comprises one or more signals OS indicative for the pump controller state and/or one or more signals indicative for the supplied one or more simulated pump control signals PC. In the example shown in FIG. 7A, the one or more simulated pump control signals PC are supplied to the neural network processor in combination with the simulated sensor signals PS. In the example of FIG. 7B, both the one or more signals OS indicative for the pump controller state and the one or more signals indicative for the supplied one or more simulated pump control signals PC are supplied to the neural network processor in combination with the simulated sensor signals PS. In the example of FIG. 7C, only the one or more signals OS indicative for the pump controller state are supplied to the neural network processor in combination with the simulated sensor signals PS.

[0065] In step S5, the neural network processor computes one or more output signals NO in response to the supplied combination of input signals. The combination of input signals comprises at least the one or more simulated sensor signals. In addition the combination of input signals comprises either the one or more signals OS indicative for the pump controller state or the one or more simulated pump control signals PC or the pump controller state signals OS and the simulated pump control signals PC. As noted the neural network processor to be trained may be either a separate computational unit, or may be part of a common computational unit. The one or more output signals NO are a tentative diagnostic indication for the condition of the simulated controlled liquid chromatography pump unit. During the execution of the training method the quality of the output signals as a diagnostic indication gradually improves.

[0066] In step S6 a loss function is computed by comparison of a diagnostic state as indicated by the one or more output signals NO and a diagnostic state as indicated by the pump fault simulation signal PFC.

[0067] In step S7 the neural network processor is trained by feeding back a loss computed with the loss function. This can for example be achieved by a back propagation computation.

[0068] The steps S1-S7 can be repeated one or more times for various conditions indicated by the pump fault simulation signal PFC. Typically for each condition indicated by the pump fault simulation signal the sequence of steps is performed a plurality of times, therewith introducing noise in the input signals of the neural network processor to avoid over fitting. Also the repetitions will typically include a plurality of sequences for the case that the pump fault simulation signal PFC indicates the absence of a pump fault. The repetitions for a pump fault simulation signal PFC indicating a particular pump fault, combination of pump faults or absence of pump faults do not need to immediately succeed each other, but simulations for various conditions may be alternated.

[0069] The method can be discontinued once it is known, or can be assumed that the neural network processor is sufficiently trained to function as a trained neural network processor 7 in a controlled liquid chromatography pump unit 1 with auto-diagnostic facility, as shown in FIG. 1, 2, or in a diagnostic unit 70 for use in a system with a controlled liquid chromatography pump unit that does not require its own diagnostic facility, for example see 1A, 1B in FIGS. 3, and 4. Upon detecting that the loss computed in step S6 is sufficiently low, it is ascertained that the neural network processor is sufficiently trained for this purpose. Alternatively, it may be presumed that the neural network processor is sufficiently trained if the sequence of steps S1-S7 was repeated for at least a predetermined minimum number of times for each condition to be detected. In case the neural network processor is considered to be sufficiently trained its parameters can be frozen. In this state the neural network processor is named the trained neural network processor. If desired the trained neural network processor can be cloned for example to have a respective trained neural network processor 7 available for auto-diagnostic purposes in a controlled liquid chromatography pump unit 1 as shown in FIGS. 1, and 2, or to be incorporated in respective diagnostic units 70 for use in a system as shown in FIG. 3 or 4. Alternatively, it may be contemplated to directly use the trained neural network processor 7 obtained with the method of FIGS. 6 and 7A, 7B or 7C. It may even be the case that a general purpose computation unit which is used for performing the method of FIGS. 6 and 7A, 7B or 7C is also used to execute the trained neural network processor 7 on the general purpose computation unit. For example a remote service may have a diagnostic unit interface 71 to enable the trained neural network processor 7 on the general purpose computation unit to receive the combination of input signals from a client as shown in FIGS. 3 and 4 respectively.

[0070] FIG. 8A shows an exemplary neural network processor module 7s as employed in FIG. 7A, 7B, 7C in more detail. As shown therein the exemplary neural network processor module 7s comprises a trainable neural network 7a that receives a combination of input signals comprising at least the simulated sensor signals PS and furthermore at least one of simulated pump control signals PC and signals OS indicative for an operational state in the pump controller simulation module 2s. In response thereto, the trainable neural network processor 7A computes S5 an output signal NO which is a tentative diagnostic signal indicative for the fault which is being simulated as specified by the pump fault simulation signal PFC. As indicated in FIG. 9A by the dashed arrow, the output signal NO cannot yet be used for actual diagnostic purposes. However, the output signal NO is provided to a loss computation unit 71 that compares the tentative diagnosis indicated by the signal NO with a Ground Truth for the diagnosis, as specified by the pump fault simulation signal PFC and in response thereto computes S6 a loss signal LS, that is indicative for a magnitude of the deviation between the tentative diagnosis and the Ground Truth. In step S7, the loss signal LS is used to train the trainable neural network processor 7a. This for example achieved by back propagation.

[0071] After a sufficient number of cycles a trained neural network processor 7b is obtained as shown in FIG. 8B. In this stage the loss computation unit 71 is no longer required, and the trained neural network processor 7b is fit for use in a diagnostic module of a controlled liquid chromatography pump unit 1 with auto-diagnostic facility or for use as a separate diagnostic unit in a system with one or more controlled liquid chromatography pump units 1.

[0072] In one exemplary embodiment, as shown in FIG. 9, the neural network processor used is a Multilayer Perceptron (MLP) 720 cooperating with a feature extraction component 710 that preprocesses the combination of input signals. The feature extraction component 710 for example computes the following signals. In response to a primary pressure signal S64 from pressure sensor 64 it computes a minimum value, a mean value and a maximum value for the pressure indicated by the signal S64 during a time window. Similarly, it computes a minimum value, a mean value and a maximum value for the pressure indicated by the signal S67 during a time window. In the embodiment shown it computes similar statistical data for the encoder signals S49, S59, indicative for the position of the pistons. It may further compute similar statistical data for a flow in case flow sensors 63, 66, 68 are provided. In the example of FIG. 9, the Multilayer Perceptron (MLP) 720 comprises an input layer to receive the preprocessed signals, a first, a second and a third dense layer and an output layer. FIG. 9A schematically shows an example of a neural network computation element 721. It comprises a transfer function which adds the weighted inputs x1, . . . ,xn and supplies the weighted sum so obtained to a non-linear activation function, here a threshold function is used which issues a binary 1 if the weighted sum exceeds a threshold and a binary 0 otherwise. The MLP 720 was trained with an Adam optimizer using default parameters (learning rate=0.001, beta 1=0.9, beta 2=0.999 and no AMSGrad).

[0073] FIG. 10 shows a second exemplary embodiment wherein the neural network processor 7 which comprises amongst others a combination of Convolutional Neural Networks (CNN) 740 and a pair 760 of Long Short-Term Memory (LSTM) layers. In the example shown in FIG. 10, the inputs of the input layer 730 comprise recordings of the sensors signals and the pump controller signals. In this example 100 samples at mutually subsequent points in time during a time window are obtained of each input signal). CNN's 740 are used to abstract the inputs into “features” stored in the feature maps 1, 2, . . . m to therewith provide a more dense representation of the original input signals. A Max-Pooling layer 750 is employed to provide these features as an input for the plurality of filters formed by LSTM layers 760. LSTM layers, unlike standard feedforward neural networks, have feedback loops, which make them more suited to translate the feature sets into the specific faults in an UHPLC pump. The outputs of the LSTM layers 760 is provided to a dense layer 770 and an output layer 780 to compute the diagnostic output NO (7DO).

[0074] An example of an LSTM can be seen in FIG. 10A. Compared with standard region based neural networks (RNNs), LSTMs are capable of learning long-term dependencies. In the example of FIG. 10A, this is achieved by using four interacting neural network layers (σ, σ, tanh and σ) instead of only one layer. The four interacting neural network layers together determine what input information should be kept and how much it should contribute to the output. The exemplary LSTM shown in FIG. 10A further has three inputs (X.sub.t−1, X.sub.t, X.sub.t+1), three outputs (h.sub.t−1, h.sub.t, h.sub.t+1) and multiple LSTM cells A. More specifics are described in Christopher Olah, ed. (2019), “Understanding LSTM Networks”, https://colah.github.io/posts/2015-08-Understanding-LSTMs/.

[0075] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and/or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. Further, unless expressly stated to the contrary, “or” refers to an inclusive or and not to an exclusive or. For example, a condition A or B is satisfied by any one of the following: A is true (or present) and B is false (or not present), A is false (or not present) and B is true (or present), and both A and B are true (or present).